Compositional Kronecker Context Optimization for vision-language models

Context Optimization (CoOp) has emerged as a simple yet effective technique for adapting CLIP-like vision- language models to downstream image recognition tasks. Nevertheless, learning context with satisfactory base-to-new, domain and cross-task generalization ability simultaneously while adapting to new tasks is a challenge. To tackle such a challenge, existing methods mainly exploit knowledge distillation with auxiliary text data written by human experts. However, we instead explore a new technique route by structuring the prompts without resorting to extra text data. As a result, we obtain a new lightweight yet generalizable approach termed Compositional Kronecker Context Optimization (CK-CoOp). Technically, the prompt's context words in CKCoOp are learnable vectors, which are crafted by linearly combining base vectors from a dictionary. These base vectors consist of a non-learnable component obtained by quantizing the weights in the token embedding layer, and a learnable component constructed by applying Kronecker product on several learnable tiny matrices. Intuitively, the compositional structure mitigates the risk of overfitting on training data and the Kronecker product breaks the non-learnable restrictions of the dictionary, thereby enhancing representation ability with minimal additional parameters. Extensive experiments confirm that, compared with existing methods, CKCoOp can not only achieve comparable or even better performance under base-to-new, domain and cross-task generalization evaluation without the help of auxiliary text data, but also has the merits of fewer learnable parameters and efficient training and inference speed.